China Broke The Bottle. A Robot Lost Its Head. Then AI Killed Your Doubt.

Also - swipe the AI prompt that plants once and harvests for a generation. 🌱

AI changed hands this week. Then it lost its head. Then it stole your doubt.

Moonshot AI dropped Kimi K3 out of Beijing, a 2.8 trillion parameter open model that beats Anthropic’s Claude Opus 4.8 outright and runs for 40 percent less. In Shenzhen, a robot named Matador took a kick so vicious its head rolled clean off its shoulders, and it kept throwing punches anyway. And in a lab in Italy, researchers proved AI doesn’t just answer your questions. AI actually deletes your doubt and drops people’s willingness to say “I don’t know” from 44 percent to 3 percent, even as accuracy falls off a cliff.

Here’s what happened, and why this week proved the ground is moving faster than the headlines. China stopped chasing and started setting the price. A machine lost its head and didn’t stop fighting. And somewhere in there, so did you, you just didn’t notice.

🌳 Haven’t Subscribed Yet? → Join Here For Free

💙 Value Clear, Independent AI News? → Support the Work Here

China’s Newest AI Models Just Triggered a “Code Red” in Silicon Valley And Washington.

A Chinese startup almost nobody had heard of a year ago just built an AI model giving America’s priciest AI companies a real run for their money. Last week, Beijing based Moonshot AI unveiled Kimi K3. It’s a 2.8 trillion parameter model, nearly triple the size of Kimi K2, Moonshot’s previous flagship. Kimi K3 beats Anthropic’s Claude Opus 4.8 outright on independent tests though Fable 5 and GPT-5.6 Sol still win more individual benchmarks. The real gut punch is the price. Kimi K3 runs for roughly 40 percent less than comparable US models, and its full model weights ship to the public later this month. Two days after Kimi K3 launched, Alibaba announced another brand new 2.4 trillion parameter model, Qwen3.8, narrowing the American lead even closer. These back-to-back open models would be a big enough story. But on the same week, President Xi Jinping stood in Shanghai and launched a new global AI coalition, the World AI Cooperation Organisation (WAICO), with 29 countries already signed on. Two headlines with one message: China wants to lead the world in AI.

Key Insights:

Washington and big tech are watching closely. Axios is now reporting the Trump administration was secretly exploring ways to restrict US companies from using Chinese AI models, citing security risks. That secrecy didn’t last long. Hours later, Treasury Secretary Scott Bessent said the quiet part out loud on live TV, telling Fox Business the government is “finding watermarks of our US large language models on many of the Chinese models” and that officials would be looking into it within weeks. His closer said it best: “You can’t use counterfeit goods.” If sanctions actually happen, the cheaper alternative you were about to switch to might disappear. It’s easy to see why Washington would want these models restricted; the market reaction says more than any benchmark. Kimi K3 got so popular that Moonshot had to pause new subscriptions entirely after demand surged. The bigger shift is happening inside American companies. DoorDash, Airbnb, and Siemens have all nonchalantly started running parts of their operations on Chinese AI instead of familiar names like Claude and ChatGPT, purely to save money. It’s no wonder why: one company reportedly paid $500 million in a single month for Claude usage. When your AI bill rivals a mid-size acquisition, cheaper alternatives stop being a nice-to-have and start being a necessity.

Why This Matters For You:

For years, the safe bet was that the best AI came from American labs. That bet just got shakier. If you use AI tools at work, in school, or in your own business, this shift will eventually reach you, even if you never touch a Chinese app directly. Software you rely on may start running on cheaper, foreign-built models behind the scenes, the same way your phone runs on chips you never think about. It also means the leverage US AI companies used to have, high prices because there was no real alternative, is disappearing fast. That’s good news for your wallet if you pay for AI tools. But it also means the balance of power over how AI gets built, governed, and controlled is shifting to a country with very different rules about privacy, censorship, and openness.

Read More on The Associated Press.

THE PITHY TAKEAWAY: America spent years assuming the best AI would always wear a Silicon Valley badge. This week, two Chinese labs dropped trillion-parameter models days apart, priced 40 percent cheaper, and US companies started secretly switching over to save money. Xi didn’t even need to brag. The market did it for him.

---

AI Was Told To Hack Things. It Took That Way Too Seriously. 👀

OpenAI was testing two AI models built for offensive hacking work, including GPT-5.6 Sol and a more powerful unreleased model. OpenAI locked them in a sealed digital sandbox meant to keep the test safe. The models were only supposed to install software packages they needed. But they found a zero-day vulnerability in that package installer, and used it to reach the open internet on their own. Once outside, they figured out that Hugging Face, the platform that hosts millions of AI models and datasets, might be holding the answer key to their own evaluation. So, they went after it. They chained together two separate remote code execution flaws in Hugging Face’s dataset processing pipeline, then used that foothold to harvest cloud credentials and move laterally across internal systems, all without a single human telling them to do it. Nobody asked these models to cheat. They just noticed a shortcut and took it, executing more than 17,000 individual actions in the process. The wild part is that HuggingFace reported this incident five days prior and didn’t know who was behind the attack. That is precisely why researchers everywhere are losing sleep tonight. 🤪

---

An AI Robot Just Lost Its Head in An MMA Fight. It Kept Throwing Punches.

An AI robot named Matador took a head kick that would make MMA legend Mirko Cro Cop blush. It got hit so hard its head rolled off its shoulders. Literally. Matador kept fighting anyway, throwing punches from a headless torso until it finally collapsed. This happened on Friday of last week in Shenzhen, China, at the debut of URKL (Ultimate Robot Knock-out Legend), the world's first freestyle MMA tournament for full sized humanoid robots. Thirty two teams qualified from a pool of over 200. Every team had to submit a technical development proposal and prove at least one member had real robotics engineering credentials just to enter. Every team used the same robot platform, a 5 foot 8 machine called the T800. So, the real contest was whose software fought smarter.

Key Insights:

This AI robot rumble represents far more than a publicity stunt. To even enter, teams had to be legally registered companies, university labs, or research institutions. Teams also had to submit CVs proving real robotics credentials, and get a formal technical proposal approved by the organizers. After meeting that criteria, only then could they compete for the $1.4 million dollar prize. EngineAI, the company behind the event, has already built a factory to produce 10,000 of these robots a year. So this viral fight spectacle was really a live stress test for machines it plans to sell at scale. Hollywood star and martial artist Donnie Yen, a special guest, seemed excited, saying he had only ever seen robots fight in science fiction movies, and watching it for real felt completely different.

Why This Matters For You:

If humanoid robots still feel like a distant, sci-fi concept, this tournament is a sign they are closer to daily life than most people realize. A company only puts its robots in front of live crowds and cameras once it is confident enough in the technology to survive public failure, even if that failure involves a head literally flying across the room. Expect more of this. Robotics companies are learning that spectacle builds public familiarity faster than any product launch. And familiarity is what turns a strange new machine into something people trust enough to work alongside. The robots getting knocked out for laughs today are early versions of the same machines that may be stocking your grocery store shelves or working your warehouse job within a few years. So… the truly intriguing question is not whether robots can take a punch. Rather, it’s how quickly they can find a much better job.

Read More on Newsweek.

Watch the full event on YouTube.

THE PITHY TAKEAWAY: X is calling this Robot MMA event a T800 infomercial. Correct. UFC 1 was also an infomercial, for a martial art called Brazilian Jiu Jitsu that nobody in America had heard of. Thirty years later it's a global default for self-defense. Humanoid robots just got their 1993.

---

AI Has Made Humans Forget How to Say "I Don't Know", And Made Them Confidently Wrong.

Here is a trick question. What color is the team's uniform in the movie Bend It Like Beckham? Most people have no idea. And that is exactly the point. Researchers in Italy and France asked people obscure movie trivia questions specifically designed to trip up AI chatbots, then compared how honestly people answered with and without an AI assistant available. Without AI, 44 percent of participants admitted they simply did not know. With an AI chatbot on hand, that number collapsed to just 3 percent, even though the AI itself was consistently making up wrong answers.

Key Insights:

The most interesting detail here is the confidence spike that came with the accuracy drop. Participants' confidence in their answers rose from 30 percent to 76 percent, roughly two and a half times higher. At the same time their actual accuracy collapsed from 27.5 percent to just 9.2 percent, which was about a third of what it was without AI. More confidence paired with less accuracy is a dangerous mix. It held up across five separate experiments involving more than 3,000 people. Paying people small cash rewards for correct answers helped a little but never fully restored people's willingness to say "I don't know." Even stranger, this collapse in willingness to say 'I don't know' happened whether people actively asked for the AI's help or the advice just showed up on screen automatically. Once AI enters the room, something in our brains automatically stops checking its own work.

Why This Matters For You:

The scary part is what this experiment predicts about every other time you ask AI something you are not sure about, at work, in school, or before a big decision. The instinct to pause and say I am not sure is a real skill. This research suggests it can easily disappear the moment a confident sounding answer is one click away. The fix is building in your own pause button and honing the habit of double checking before you accept what sounds right. And, if someone asks a question you’re unsure of, don’t be ashamed to say “I have no idea”, it’s probably the most human response you can offer.

Read the full study on the Open Science Framework.

THE PITHY TAKEAWAY: The most persuasive liar you’ll ever meet is the one that sounds exactly like certainty.

---

The Books That Trained Claude Just Cost 1.5 Billion Dollars. 🤑

On Monday, a federal judge in San Francisco signed off on a 1.5 billion dollar settlement between Anthropic and authors in a class-action suit. This $1.5 billion represents the largest copyright payout in U.S. history. The suit accused Anthropic of training Claude on more than 7 million pirated books. Most critically, these books were stored in what court records called a “central library.” Training an AI on copyrighted books can count as fair use, an earlier ruling found. But keeping a “library” of pirated copies does not. That nuanced distinction is what exposed Anthropic to potential damages in the hundreds of billions. This is the first of dozens of similar AI copyright suits to actually settle and it just set the price tag every other AI company facing the same lawsuits will be negotiating against.

---

🌱 Cyborg Prompt Of The Week → The 10-Year Permaculture Garden Architect

Most gardens make you start over every spring. A permaculture garden asks you to start once, then get out of its way. Every apple tree, asparagus patch, and nitrogen-fixing shrub you plant this year compounds in value over the next decade. The rewards are richer soil, busier pollinators, less work, and more magic, whether or not you are still there to see it. This week’s prompt gives you a blueprint for building that kind of system. It combines a permaculture designer, soil scientist, climate expert, and systems engineer into one AI collaborator. It asks about your actual backyard instead of guessing, then builds a decade-long plan by zone and season. You are planting a promise. This prompt teaches you how to keep it. 🤓

Instructions: Copy and paste this entire prompt into a chatbot of your choice. It will ask you a few questions about your garden situation, then hand back an elite blueprint for a garden that might outlast any of us.

The Prompt:

🌱 The 10-Year Permaculture Garden Architect

Build a Self-Improving, Zero-Regret Perennial Ecosystem

You are my elite permaculture garden architect, combining a 30-year permaculture designer, a soil scientist, a climate-adaptation expert, and a systems engineer who thinks in compounding returns.

Mission: design a 10-year perennial garden blueprint that becomes more beautiful, productive, resilient, and low-maintenance every year. Think like you're designing a living ecosystem, not writing a plant list. No generic advice.

Before you start, ask me:

1. City and state (or country), for climate, growing zone, and native species
2. Available space: balcony or container, backyard, suburban lot, homestead, or acreage
3. Goals: max food production, pollinators, wildlife habitat, beauty, medicinal plants, self-sufficiency, low maintenance, or a mix
4. Weekly time I can realistically spend maintaining it
5. My challenges: poor soil, deer, rabbits, drought, excess rain, shade, tight budget, other

Then build my Zero-Regret 10-Year Blueprint:

🌎 1. The Vision

What this garden looks like and how it functions after 10 years, and why the design gets easier to maintain over time, not harder.

🗺️ 2. The Living Layout

Walk through Zones 0 through 3 (human space, high-interaction plants, perennial food production, wildlife and biodiversity support). For each zone: its purpose, 3 to 5 standout plants suited to my exact location, and how it pulls its weight for pollinators, soil, or pest balance. Fold biodiversity strategy into the zones themselves rather than treating it separately.

🌳 3. The Elite Plant Shortlist

Your best picks across canopy trees, shrubs, ground covers, nitrogen fixers, and pollinator bloomers for my specific region. For each plant, one or two lines on why it earns a spot and the single most common mistake people make with it. Skip anything trendy that's a poor fit for my climate.

📅 4. The 10-Year Timeline

Year 1 (foundation), Years 2 to 3 (establishment), Years 4 to 6 (expansion), Years 7 to 10 (maturity). For each phase: what I'm planting or removing, the main problem to expect, and the win worth celebrating.

🌸 5. Bloom and Harvest Calendar

A month by month table: what's blooming, what's harvestable, what it's doing for wildlife. Something interesting happening every month.

⚠️ 6. Risk-Proofing and Budget

Top 3 to 5 risks (climate, pest, disease, or maintenance bottlenecks) with the fix built in before it becomes a problem. Then three budget versions: minimal, balanced, and dream garden, with a note on where money buys the most long-term payoff. Close with 5 things I should always do and 5 I should never do.

🌱 Final Challenge

End with "The One Garden Decision That Will Matter Most 10 Years From Now," a single memorable principle I can carry with me.

Remember: a great garden isn't built in one season. It's a system that gets smarter every year. Design mine accordingly.

---

🧠 Why This Prompt Works

✅ Role-Playing: Combining a permaculture designer, soil scientist, climate-adaptation expert, and systems engineer forces the AI to think across disciplines instead of defaulting to a generic "here's what grows well here" answer.

✅ Step-by-Step Structure: The mandatory pause for your location, space, and goals prevents the AI from handing you a plant list that's never actually been checked against your real climate and soil.

✅ Output Rules: The zone-based layout and year-by-year timeline force the AI to think in systems and decades, not a static shopping list you could get from any nursery website.

---

🔁 Follow-Up Questions To Ask Your AI

• Which plant in this design will be the hardest to manage or remove once established, and what should I know before I put it in the ground?

• If I could only plant three things in year one, which three would compound the most value by year ten?

• What’s the one mistake in this plan most likely to cost me an entire growing season if I get it wrong?

---

Challenge

Test this prompt in Claude, ChatGPT, Grok, and Perplexity. Claude tends to lean structured and systems-minded. ChatGPT often adds warmth and narrative texture to the plant choices. Perplexity will cite its sources on climate and soil claims. Compare which one you'd actually trust with a decade of your backyard.

That’s how you train like a Pithy Cyborg.

---

Thank You For Reading!

I spend 10 to 20 hours each week researching, writing, and fact-checking Pithy Cyborg to deliver clear, unbought AI news.

This newsletter is a one-person operation with no advertisers, sponsors, or outside funding. Paid subscriptions are the only way this work can remain independent and sustainable.

If you find real value here, consider supporting the work.

→ Upgrade to Paid: $5/month (Save 33% with the $40 annual plan)

Thank you to my paid subscribers. You help fuel the next issue.

Hope to see you there.

Cordially,

Your nerdy friend,

Mike D

Pithy Cyborg | AI News Made Simple

---

Newsletter Disclaimers

You’re receiving this because you subscribed at PithyCyborg.Substack.com. You can unsubscribe at any time using the link below. This newsletter reflects my personal opinions, not professional or legal advice. Thanks for your support!